Aircraft maintenance stand scheduling method based on heterogeneous graph model

Abstract

The invention relates to an aircraft maintenance position scheduling method based on a heterogeneous graph model, and relates to the technical field of aircraft maintenance intelligent scheduling. According to the scheme, input feature vectors including maintenance machine position feature vectors, maintenance item feature vectors and O-M arc feature vectors are obtained, and a heterogeneous graph model including maintenance machine position nodes and maintenance item nodes is constructed based on the input feature vectors; and taking the input feature vector and the heterogeneous graph model as the input of the aircraft maintenance stand scheduling strategy prediction network to realize the prediction of the aircraft maintenance stand scheduling strategy. The aircraft maintenance stand scheduling strategy prediction network is realized based on a graph neural network, features of each type of nodes are sufficiently and independently extracted, interference caused by feature mixing is avoided, and meanwhile, mutual influence of the two types of nodes is reserved through information transmission between the two types of nodes. The accurate expression of node features is effectively improved, and the flexibility and efficiency in processing a complex scheduling scene are remarkably enhanced, so that a finer scheduling decision is obtained.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent scheduling of aircraft maintenance, and in particular to an aircraft maintenance stand scheduling method based on a heterogeneous graph model. Background Art

[0002] The maintenance areas of airlines and large maintenance bases are distributed with maintenance stands with different maintenance capabilities. From the time an aircraft to be maintained enters the maintenance area to the time it leaves the maintenance area, it needs to complete all the maintenance items it carries on different maintenance stands. However, in order to improve the space utilization rate of the maintenance area, civil aviation maintenance companies place aircraft in the maintenance area densely, and the distance between aircraft is reduced, resulting in a sharp increase in the risk of collisions and scratches caused by aircraft replacement maintenance stands. At present, the domestic maintenance stand resource scheduling generally adopts manual methods or traditional operations research methods. With the increase in maintenance aircraft and maintenance tasks, the workload increases exponentially, the resource scheduling effect is poor, and the safety risk factor decreases. Therefore, the reasonable scheduling of maintenance stands is crucial, which can improve the safety of aircraft maintenance, reduce the loss and cost of aircraft collisions, and improve maintenance production efficiency, which is of great significance to the guarantee of airport operation and maintenance resources. Summary of the invention

[0003] The technical problem to be solved by the present application is to provide an aircraft maintenance stand scheduling method based on a heterogeneous graph model, which has the characteristics of improving the rationality of maintenance stand scheduling and being easy to implement in engineering.

[0004] In one embodiment, a method for scheduling aircraft maintenance stands based on a heterogeneous graph model is provided, comprising: Obtain an input feature vector; the input feature vector includes a maintenance station feature vector, a maintenance project feature vector and an OM arc feature vector; wherein the maintenance station feature vector includes: the earliest available time of each maintenance station, the number of maintenance projects that can be repaired by each maintenance station in the current maintenance project, and the utilization rate of each maintenance station; the maintenance project feature vector includes: the maintenance completion status of each maintenance project, the number of available maintenance stations for each maintenance project, the average maintenance time of each maintenance project, the number of maintenance projects that have not started maintenance, and the start maintenance time of each maintenance project; the OM arc feature vector includes the maintenance time of each maintenance project at each available maintenance station; wherein the utilization rate of any maintenance station represents the ratio of the time difference between the start maintenance time of the maintenance project being repaired at any maintenance station and the current time to the time required for the maintenance project being repaired; Constructing a heterogeneous graph model based on the input feature vector, the heterogeneous graph model includes a maintenance stand node and a maintenance item node, wherein, for each maintenance aircraft, a connection arc between maintenance items is constructed according to the dependency relationship between various maintenance items of the maintenance aircraft, and for each maintenance stand, an OM arc is established according to the maintenance items that can be maintained at each maintenance stand; The input feature vector and the heterogeneous graph model are used as inputs of an aircraft maintenance stand scheduling strategy prediction network to achieve prediction of the aircraft maintenance stand scheduling strategy; wherein the aircraft maintenance stand scheduling strategy prediction network is obtained by learning and training in combination with a constructed loss function; Scheduling of aircraft maintenance stands is achieved based on the predicted aircraft maintenance stand scheduling strategy; The method of using the input feature vector and the heterogeneous graph model as inputs of an aircraft maintenance stand scheduling strategy prediction network to achieve prediction of an aircraft maintenance stand scheduling strategy includes: The original feature vector and the heterogeneous graph model are used as inputs. Based on the first graph attention network, through weighted aggregation, the maintenance project node aggregates the information of the neighboring maintenance project nodes, and the maintenance project node embedding vector is calculated. The original feature vector and the heterogeneous graph model are used as input. Based on the second graph attention network, the maintenance station node aggregates the information of the surrounding neighbor nodes through weighted aggregation, and the maintenance station node embedding vector is calculated. Performing a flat pooling operation on the maintenance position node embedding vector and the maintenance item node embedding vector to obtain a state embedding vector; The maintenance position node embedding vector, the maintenance item node embedding vector and the state embedding vector are used as inputs of a multilayer perceptron, and the action scalar of each executable OM arc is obtained through processing by the multilayer perceptron; Each action scalar is normalized to obtain the aircraft maintenance stand scheduling strategy.

[0005] The beneficial effects of the present invention are: In the aircraft maintenance stand scheduling method based on heterogeneous graph model, on the one hand, based on the first graph attention network, through weighted aggregation, the maintenance project node aggregates the information of neighboring maintenance project nodes, and the maintenance project node embedding vector is calculated. On the other hand, based on the second graph attention network, through weighted aggregation, the maintenance stand node aggregates the information of surrounding neighboring nodes, and the maintenance stand node embedding vector is calculated. Then, the maintenance stand node embedding vector and the maintenance project node embedding vector are subjected to a flat pooling operation to obtain the state embedding vector, so that the maintenance stand node embedding vector, the maintenance project node embedding vector and the state embedding vector can be used as the input of the multi-layer perceptron, and the action scalar of each executable OM arc obtained by the multi-layer perceptron processing. Therefore, each type of node is fully extracted separately to avoid interference caused by feature mixing, and at the same time, the mutual influence of the two types of nodes is retained through information transmission between each other. It not only effectively improves the precise expression of node features, but also significantly enhances the flexibility and efficiency in dealing with complex scheduling scenarios, so as to obtain more refined scheduling decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0006] Figure 1 This is a flow chart of a training method for an aircraft maintenance stand scheduling strategy prediction network according to an embodiment of the present application; Figure 2 This is a schematic diagram of a network training architecture for predicting aircraft maintenance stand scheduling strategies according to an embodiment of the present application; Figure 3 is a schematic diagram of a heterogeneous graph model according to an embodiment of the present application; Figure 4 This application Figure 1 A method flow diagram of an embodiment of step S30; Figure 5 This application Figure 4 A method flow diagram of an embodiment of step S301; Figure 6 This application Figure 4 A method flow diagram of an embodiment of step S302; Figure 7 It is a schematic diagram of a flow chart of an aircraft maintenance stand scheduling method based on a heterogeneous graph model according to an embodiment of the present application; Figure 8 This application Figure 7 A method flow chart of an embodiment of step S300 in FIG. DETAILED DESCRIPTION

[0007] The present invention is further described in detail below by specific embodiments in conjunction with the accompanying drawings. Wherein similar elements in different embodiments adopt associated similar element numbers. In the following embodiments, many detailed descriptions are for making the present application better understood. However, those skilled in the art can easily recognize that some features can be omitted in different situations, or can be replaced by other elements, materials, methods. In some cases, some operations related to the present application are not shown or described in the specification, this is to avoid the core part of the present application being overwhelmed by too much description, and for those skilled in the art, it is not necessary to describe these related operations in detail, and they can fully understand the related operations according to the description in the specification and the general technical knowledge in the art.

[0008] In addition, the features, operations or characteristics described in the specification can be combined in any appropriate manner to form various implementations. At the same time, the steps or actions in the method description can also be interchanged or adjusted in a manner that is obvious to those skilled in the art. Therefore, the various sequences in the specification and the drawings are only for the purpose of clearly describing a certain embodiment and are not meant to be a required sequence, unless otherwise specified that a certain sequence must be followed.

[0009] The serial numbers assigned to the components in this article, such as "first", "second", etc., are only used to distinguish the objects described and do not have any order or technical meaning.

[0010] To facilitate the description of the inventive concept of the present application, the aircraft maintenance stand scheduling technology is briefly described below.

[0011] In the current aircraft maintenance scheduling technology, in one technology, the data of aircraft to be repaired and the maintenance task data of the aircraft to be repaired are collected. Based on the collected data, the number of movements of all aircraft to be repaired is minimized as the optimization goal, constraints are set, and iterative solutions are performed to achieve maintenance stand scheduling of aircraft to be repaired based on the obtained maintenance stand arrangement, which can effectively improve the utilization rate of stand resources.

[0012] With the development of neural networks based on deep learning, the use of graph neural networks to implement predictions for certain technical solutions can have stronger robustness and memory capabilities.

[0013] In view of this, an embodiment of the application provides an aircraft maintenance stand scheduling method based on a heterogeneous graph model. In this scheme, an input feature vector including a maintenance stand feature vector, a maintenance item feature vector, and an OM arc feature vector is obtained, and a heterogeneous graph model including maintenance stand nodes and maintenance item nodes is constructed based on the input feature vector. Then, the input feature vector and the heterogeneous graph model are used as inputs of an aircraft maintenance stand scheduling strategy prediction network to achieve the prediction of aircraft maintenance stand scheduling strategy. Among them, the aircraft maintenance stand scheduling strategy prediction network is implemented based on a graph neural network, so that it has higher robustness while improving the rationality of maintenance stand scheduling.

[0014] However, the applicant discovered during the research that since the heterogeneous graph model obtained includes maintenance position nodes and maintenance project nodes, if the mixed embedding method of node feature vectors in the current technology is used, it will cause interference and fail to fully extract the features of the mutual influence of the two types of nodes, resulting in low accuracy in the expression of node features in the scheduling plan, reducing the flexibility and efficiency when dealing with complex scheduling scenarios.

[0015] In view of this, in the embodiment of the present application using the aircraft maintenance stand scheduling strategy prediction network to realize the aircraft maintenance stand scheduling strategy prediction, the original feature vector and the heterogeneous graph model are used as inputs. On the one hand, based on the first graph attention network, through weighted aggregation, the maintenance project node aggregates the information of the neighboring maintenance project nodes, and the maintenance project node embedding vector is calculated. On the other hand, based on the second graph attention network, through weighted aggregation, the maintenance stand node aggregates the information of the surrounding neighboring nodes, and the maintenance stand node embedding vector is calculated, and then the maintenance stand node embedding vector and the maintenance project node embedding vector are subjected to a flat pooling operation to obtain a state embedding vector, so that the maintenance stand node embedding vector, the maintenance project node embedding vector and the state embedding vector can be used as the input of the multi-layer perceptron, and the action scalar of each executable OM arc obtained by the processing of the multi-layer perceptron. In this way, each type of node is fully extracted separately to avoid interference caused by feature mixing, and at the same time, the mutual influence of the two types of nodes is retained through information transmission between each other. Not only the accurate expression of node features is effectively improved, but also the flexibility and efficiency in processing complex scheduling scenarios are significantly enhanced to obtain more refined scheduling decisions.

[0016] In order to fully illustrate the aircraft maintenance stand scheduling method based on the heterogeneous graph model in the embodiment of the present application, the training method of the aircraft maintenance stand scheduling strategy prediction network is first described.

[0017] In one embodiment, please refer to Figure 1 and Figure 2 The training method of the aircraft maintenance stand scheduling strategy prediction network of the present application may include: Step S10, obtaining the trained original feature vector, which includes a maintenance position feature vector, a maintenance item feature vector and an OM arc feature vector.

[0018] Based on the real aircraft maintenance stand scheduling data, the original feature vector of training is obtained to provide data input for the subsequent node embedding process and training process.

[0019] Among them, the maintenance position feature vector Includes: The earliest available time t for each maintenance stand k , the number of maintenance items that can be maintained at each maintenance station among the current maintenance items N (M k ) and the utilization rate of each maintenance station. For any maintenance station, since it may be in an occupied state due to the maintenance task at the current moment, the earliest available time of the maintenance station is the completion time of the current maintenance task of the maintenance station. For the items to be repaired, not all maintenance items can be repaired in any maintenance station. Therefore, it is necessary to obtain the number of maintenance items with maintenance capabilities among the current items to be repaired for each maintenance station. For the utilization rate of each maintenance station, the utilization rate of any maintenance station represents the ratio of the time difference between the start time of the maintenance item being repaired at any maintenance station and the current time to the time required for the maintenance item being repaired, which can be expressed as: ; in, Indicates the current moment, Indicates the start time of the maintenance item currently being maintained, where ij represents the jth maintenance item of the i-th aircraft, k represents the k-th maintenance stand, Indicates the duration of the maintenance project currently under repair.

[0020] Maintenance item feature vector Includes: Repair completion status of each repair item , the number of maintenance bays available for each maintenance project N (M ijk ), average maintenance time for each maintenance item , the number of maintenance items that have not started maintenance and the start maintenance time of each maintenance item. Among them, the maintenance completion status can include two states: unfinished and completed. In some embodiments, the unfinished state can be represented by 0 and the completed state can be represented by 1. For any maintenance item, not all maintenance positions have the ability to repair any maintenance item. Therefore, we can obtain the number of maintenance positions available for any maintenance item based on the maintenance capacity of each maintenance position. If a maintenance item can be repaired at any one of more than two maintenance positions, the maintenance time when each maintenance position is selected can be obtained, so that the average maintenance time can be obtained, and thus the average maintenance time for each maintenance item can be obtained.

[0021] It should be noted that, in the training phase of the network model, since the original feature vectors for training are obtained based on real aircraft maintenance stand scheduling data, the start maintenance time of each maintenance item is known, but in the prediction phase, the start maintenance time is not always known. Therefore, in the prediction phase using the trained aircraft maintenance stand scheduling strategy prediction network, the start maintenance time of each maintenance item is an estimated value. In one embodiment, the estimated value is obtained based on the latest estimated start maintenance time of each maintenance item.

[0022] OM arc eigenvector Includes the maintenance time required for each maintenance item at each available maintenance bay.

[0023] Step S20, constructing a heterogeneous graph model based on the original feature vector.

[0024] The heterogeneous graph model includes maintenance stand nodes and maintenance item nodes. For each maintenance aircraft, the connection arcs between the maintenance items are constructed according to the dependency relationship between the maintenance items of the maintenance aircraft. For each maintenance stand, the OM arc is established according to the maintenance items that can be maintained at each maintenance stand.

[0025] In one embodiment, please refer to Figure 3 , Figure 3 In the maintenance machine nodes, M1, M2, M3 and M4 are included, and the maintenance project nodes include O 11 , O 12 , O 13 , O 21 , O 22 , O 31 , O 32 , O 33 , O 41 , O 42 , and O 43 Here, M kTo represent different maintenance positions, O ij To represent different maintenance project nodes. The maintenance position node and the matching maintenance project node are represented by OM arcs, such as Figure 3 The middle dashed line, OM arc is a non-connected arc, which represents the matching relationship and competition relationship between the maintenance station and the maintenance project, thus obtaining an unscheduled heterogeneous graph model. There are two types of neighbor nodes of the maintenance project, namely, neighbor maintenance project nodes (with connecting arcs between them, Figure 3 The solid line in the middle connects the nodes and the neighboring maintenance station nodes. The neighboring nodes of the maintenance station have only one type, namely the neighboring maintenance project node.

[0026] At the same time, considering that the maintenance items of the same aircraft must be processed in chronological order, this condition is used as a constraint to establish a heterogeneous graph model. For example, for an aircraft with i=1, there are three maintenance items, namely maintenance item 11 with j=1, maintenance item 12 with j=2, and maintenance item 13 with j=3. There is a processing order for the three maintenance items, namely maintenance item 11, maintenance item 12, and maintenance item 13.

[0027] In this way, through the above-mentioned node relationship expression, the heterogeneous graph model provided by this application can better display complex node relationships and dependencies, and achieve a more complete expression of the maintenance area.

[0028] Step S30, taking the original feature vector and the heterogeneous graph model as input, and combining the loss function, the constructed initial architecture for predicting the aircraft maintenance stand scheduling strategy is trained.

[0029] Please refer to Figure 4 , step S30 may include: Step S301, taking the original feature vector and the heterogeneous graph model as input, based on the first graph attention network, through weighted aggregation, the maintenance project node aggregates the information of the neighboring maintenance project nodes, and calculates the maintenance project node embedding vector.

[0030] For the embedding of maintenance project nodes, the original feature vector designed in the previous order is used as the input of the first graph attention network (belonging to the core functional layer), focusing on the priority constraints between maintenance projects and aggregating the information of surrounding neighbor nodes to extract features for each maintenance project node. Among them, the priority constraints include: maintenance projects of the same aircraft must be processed in chronological order.

[0031] In one embodiment, in order to better capture deep-level features, the calculation of the maintenance project node embedding vector is implemented in a multi-layer stacking manner. To implement the multi-layer stacking manner, in one embodiment, the first graph attention network includes a multi-layer graph attention network. Except for the first layer of graph attention network, for any layer of graph attention network, the output of the previous layer of graph attention network is used as the input of the any layer of graph attention network.

[0032] In one embodiment, please refer to Figure 5 For any maintenance project node, based on any layer of graph attention network, through weighted aggregation, the maintenance project node aggregates the information of neighboring maintenance project nodes, and the maintenance project node embedding vector is calculated, including: Step S3011, based on the input feature vector and the heterogeneous graph model, calculate the attention coefficients between all maintenance project nodes that have connection arcs with the arbitrary maintenance project node, wherein all maintenance project nodes that have connection arcs with the arbitrary maintenance project node include the arbitrary maintenance project node itself.

[0033] In one embodiment, step S3011 can be expressed as: ; in, Indicates that any maintenance project node O ij There is a connection arc between any maintenance project nodes O ip The attention coefficient of the l-th layer graph attention network, ij represents the j-th maintenance item of the i-th aircraft, ; LeakyReLU represents the LeakyReLU activation function; represents the weight of a single-layer feedforward neural network, and T represents transpose; Represents the calculated maintenance item node O of the l-1th layer graph attention network ij The embedding vector of , when l=1, , Represents maintenance project node O ij The eigenvector of Represents the calculated maintenance item node O of the l-1th layer graph attention network ip The embedding vector of , when l=1, , Represents maintenance project node O ip The eigenvector of express The linear transformation matrix of express The linear transformation matrix of Represents a vector merge operation.

[0034] In one embodiment, = .

[0035] Please refer to Figure 3 , for maintenance project node O 12 , all maintenance project nodes with connected arcs include O 11 , O 12 and O 13 , therefore, there is .

[0036] Step S3012, based on the normalization function, calculate the normalization coefficient of the attention coefficient.

[0037] In one embodiment, step S3012 can be expressed as: ; in, Represents the attention coefficient The normalized coefficient of Represents the self-attention coefficient The normalization coefficient of , softmax represents the normalization function, represents the set of all maintenance project nodes that have connection arcs with any maintenance project node. represents the natural exponential function.

[0038] Step S3013: Through weighted aggregation, the arbitrary maintenance project node aggregates the surrounding neighbor node information, and calculates the maintenance project node embedding vector of the arbitrary maintenance project node.

[0039] In one embodiment, step S3013 can be expressed as: ; in, Represents the maintenance item node O output by the l-th layer graph attention network ij The embedding vector of represents a non-linear activation function.

[0040] When the first graph attention network includes L layers of graph attention networks (i.e., 0≤l≤L), the output of the Lth layer of graph attention network is That is the output of the first image attention network.

[0041] Through the embodiment of step S301, the priority constraint relationship of the maintenance items is focused on, so that the embedded project node characteristics can reflect the timing and correlation between projects, further improving the rationality of the scheduling strategy.

[0042] Step S302, taking the original feature vector and the heterogeneous graph model as input, based on the second graph attention network, through weighted aggregation, the maintenance station node aggregates the surrounding neighbor node information, and calculates the maintenance station node embedding vector.

[0043] For the embedding of the maintenance station node, the original feature vector designed in the previous order is used as the input of the second graph attention network (belonging to the core functional layer).

[0044] In one embodiment, similarly, in order to better capture deep-level features, the calculation of the maintenance project node embedding vector is implemented in a multi-layer stacking manner. To implement this multi-layer stacking manner, in one embodiment, similarly, the second graph attention network includes a multi-layer graph attention network, which has the same number of layers as the first graph attention network, and except for the first layer of graph attention network, for any layer of graph attention network, the output of the previous layer of graph attention network is used as the input of the any layer of graph attention network.

[0045] In one embodiment, please refer to Figure 6 For any maintenance station node, based on any layer of graph attention network, through weighted aggregation, the maintenance station node aggregates the information of surrounding neighbor nodes, and calculates the maintenance station node embedding vector, including: Step S3021, based on the input feature vector and the heterogeneous graph model, calculate the self-attention coefficient of the arbitrary maintenance station node, and calculate the attention coefficients between all maintenance item nodes that have OM arcs with the arbitrary maintenance station node.

[0046] In one embodiment, step S3021 can be expressed as: ; ; in, Indicates that any maintenance station node M k There is an OM arc between any maintenance project nodes O ij The attention coefficient of the l-th layer graph attention network, ij represents the j-th maintenance item of the i-th aircraft; Indicates any maintenance station node M k The self-attention coefficient of the l-th layer graph attention network, k represents the k-th maintenance station; LeakyReLU represents the LeakyReLU activation function; represents the weight of a single-layer feedforward neural network, and T represents transpose; represents the calculated embedding vector of the maintenance station node of the l-1th layer graph attention network. When l=1, , represents the maintenance position feature vector; represents the extended feature vector, , Represents the calculated maintenance item node O of the l-1th layer graph attention network ij The embedding vector of , when l=1, , Represents maintenance project node O ij The eigenvector of represents the OM arc eigenvector; express The linear transformation matrix of express The linear transformation matrix of Represents a vector merge operation.

[0047] In one embodiment, = = .

[0048] Step S3022, based on the normalization function, calculate the normalization coefficient of the self-attention coefficient and the normalization coefficient of the attention coefficient.

[0049] In one embodiment, step S3022 can be expressed as: ; ; in, Represents the attention coefficient The normalized coefficient of Represents the self-attention coefficient The normalization coefficient of , softmax represents the normalization function, It represents the set of all maintenance project nodes that have OM arcs with any maintenance station node. represents the natural exponential function.

[0050] Step S3023: Through weighted aggregation, the arbitrary maintenance station node aggregates the surrounding neighbor node information, and calculates the maintenance station node embedding vector of the arbitrary maintenance station node.

[0051] In one embodiment, step S3023 can be expressed as: ; in, Represents the maintenance station node M output by the l-th layer graph attention network k The embedding vector of represents a non-linear activation function.

[0052] When the second graph attention network includes L layers of graph attention networks (i.e., 0≤l≤L), the output of the Lth layer of graph attention network is That is the output of the second graph attention network.

[0053] By processing the maintenance bay nodes individually, the static characteristics of the maintenance bay (such as occupancy status, number of repairable items, etc.) can be captured more accurately, providing more reliable input for subsequent decision-making.

[0054] This multi-layer stacking can not only extract more complex relational features, but also gradually optimize the node representation. Since the features of the OM arc remain unchanged, the original OM arc feature vector is used as the input of each layer to ensure consistency and feature transfer during the embedding process.

[0055] Through step S301 and step S302, different characteristics of node types can be captured to avoid information loss or feature confusion caused by mixed processing.

[0056] Step S303, performing a pooling operation on the maintenance station node embedding vector and the maintenance item node embedding vector to obtain a state embedding vector.

[0057] In one embodiment, step S303 can be expressed as: ; in, represents the state embedding vector, Represents the maintenance station node M output by the attention network in the second graph k The embedding vector of Represents a collection of maintenance bay nodes. Indicates the number of maintenance bay nodes. Represents the maintenance item node O output by the attention network of the first graph ij The embedding vector of Represents a collection of maintenance project nodes, Indicates the number of maintenance project nodes, Represents a vector merge operation.

[0058] Through step S303, further feature extraction is achieved, and the heterogeneous graph models of different sizes are converted into fixed-dimensional , and state embedding , which helps the network better understand the complex relationships between maintenance bays and maintenance items.

[0059] Step S304: taking the maintenance position node embedding vector, the maintenance item node embedding vector and the state embedding vector as inputs of the first multi-layer perceptron, and obtaining the action scalar of each executable OM arc through processing by the first multi-layer perceptron.

[0060] In one embodiment, step S304 can be expressed as: ; in, represents the action scalar, s represents the index of the action scalar, represents the first multi-layer perceptron, represents the OM arc eigenvector.

[0061] For example, if Figure 3 At this moment, the maintenance position nodes M1 and M2 are idle, and the maintenance project node O 12 and O 32 In the state of waiting for maintenance, the executable OM arcs include M1-O 12 、M1-O 32 、M2-O 12 and M2-O 32 There are four OM arcs, so we can calculate the corresponding four action scalars.

[0062] Step S305, normalizing each action scalar to obtain an aircraft maintenance stand scheduling strategy.

[0063] In one embodiment, step S305 can be expressed as: ; in, Represents action scalar The probability of selection, represents the natural exponential function.

[0064] Step S306, using the state embedding vector as the input of the second multi-layer perceptron, and obtaining the output after being processed by the second multi-layer perceptron as the output scalar.

[0065] In one embodiment, step S306 can be expressed as: ; Among them, v represents the output scalar, Represents the second multi-layer perceptron.

[0066] The output scalar is used as the status evaluation value of the current scheduling plan. Based on the preset objective function and combined with the loss function, the final strategy, that is, the optimal maintenance aircraft scheduling plan, is obtained through repeated iterations of the "strategy evaluation-strategy improvement" process.

[0067] Step S307, constructing a loss function based on the outputs of the first multi-layer perceptron and the second multi-layer perceptron, and back-propagating the constructed loss function to the first graph attention network and the second graph attention network to implement the training of the aircraft maintenance stand scheduling strategy prediction network.

[0068] In this process, the loss function is used to evaluate the performance of the model and guide the update of the network. Specifically, the loss function of the actor network aims to maximize the expected return of the strategy, while the loss function of the critic network minimizes the error of the state value estimation.

[0069] The critic network estimates the output scalar (value) v for a given state (the state of the maintenance bay and the maintenance item) and optimizes it by minimizing its difference from the target value. This target value is usually calculated by the actual reward r t and the discounted return of future states. Therefore, the loss function of the critic network is defined as: ; in, represents the loss function of the critic network, Indicates the status characteristics of the maintenance station and maintenance items at the current moment s t The corresponding output scalar, The state characteristics s representing the maintenance position and maintenance items at the next moment t+1 The corresponding output scalar, represents the discount factor, which is used to weigh the impact of current rewards and future rewards. Indicates the penalty imposed when an aircraft changes its maintenance stand. <0.

[0070] In one embodiment, =-1.

[0071] The goal of the actor network is to learn an action scalar (policy) , so that it chooses the action, , that is, assigning a maintenance project to a suitable maintenance station to maximize future returns. Indicates the status characteristics of the maintenance station and maintenance items at the current moment s t The corresponding selection action is Indicates selection action The loss function of the actor network is based on the policy gradient method, by introducing the advantage function To guide the strategy update, it can be expressed as: ; in, represents the loss function of the actor network, , which is used to measure how well the current action performs relative to the baseline policy.

[0072] Based on the loss function of the actor network and the loss function of the critic network, the total loss function is obtained, which can be expressed as: ; in, Represents the total loss function, c is a hyperparameter that can be set based on experience.

[0073] In one embodiment, c=0.5 may be set.

[0074] Through the above loss function, the actor-critic network can optimize the strategy and state value estimation simultaneously, so as to gradually converge to the optimal aircraft maintenance stand scheduling plan, thereby achieving the optimization goal of minimizing the number of times the aircraft changes maintenance stands.

[0075] In step S304 to step S306, a random strategy is generated based on the first multilayer perceptron, and the generated random strategy is evaluated based on the result of the second multilayer perceptron. The two multilayer perceptrons are used as the action scalar and output scalar of the action decision layer respectively, forming an action decision layer reinforcement learning algorithm of the actor-commentator framework to realize the training of the aircraft maintenance stand scheduling strategy prediction network.

[0076] Based on the aircraft maintenance stand scheduling strategy prediction network obtained through the above training, an aircraft maintenance stand scheduling method based on a heterogeneous graph model provided in one embodiment of the present application is provided. Please refer to Figure 7 ,include: Step S100, obtaining an input feature vector, which includes a maintenance station feature vector, a maintenance project feature vector and an OM arc feature vector. The maintenance station feature vector, the maintenance project feature vector and the OM arc feature vector can refer to the description of the maintenance station feature vector, the maintenance project feature vector and the OM arc feature vector in step S10.

[0077] Step S200, constructing a heterogeneous graph model based on the input feature vector.

[0078] In step S200 , the method for constructing the heterogeneous graph model may refer to the description in step S20 .

[0079] Step S300, taking the input feature vector and the heterogeneous graph model as the input of the aircraft maintenance stand scheduling strategy prediction network to realize the prediction of the aircraft maintenance stand scheduling strategy.

[0080] Please refer to Figure 8 , step S300 may include: Step S3001, taking the original feature vector and the heterogeneous graph model as input, based on the first graph attention network, through weighted aggregation, the maintenance project node aggregates the information of the neighboring maintenance project nodes, and calculates the maintenance project node embedding vector.

[0081] The specific method of step S3001 can refer to the description of step S301.

[0082] Step S3002, taking the original feature vector and the heterogeneous graph model as input, based on the second graph attention network, through weighted aggregation, the maintenance station node aggregates the surrounding neighbor node information, and calculates the maintenance station node embedding vector.

[0083] The specific method of step S3002 can refer to the description of step S302.

[0084] Step S3003, performing a pooling operation on the maintenance position node embedding vector and the maintenance item node embedding vector to obtain a state embedding vector.

[0085] The specific method of step S3003 can refer to the description of step S303.

[0086] Step S3004: taking the maintenance position node embedding vector, the maintenance item node embedding vector and the state embedding vector as inputs of the first multilayer perceptron, and obtaining the action scalar of each executable OM arc through processing by the first multilayer perceptron.

[0087] The specific method of step S3004 can refer to the description of step S304.

[0088] It should be noted here that the second multilayer perceptron is only needed in the training phase of the network. In the prediction phase based on the training, the second multilayer perceptron will not work. Therefore, in the prediction phase, the second multilayer perceptron can be removed or retained but not enabled.

[0089] Step S3005, normalizing each action scalar to obtain an aircraft maintenance stand scheduling strategy.

[0090] Step S400: Scheduling the aircraft maintenance stand based on the predicted aircraft maintenance stand scheduling strategy.

[0091] Based on the above aircraft maintenance stand scheduling method based on heterogeneous graph model, on the one hand, based on the first graph attention network, through weighted aggregation, the maintenance project node aggregates the information of neighboring maintenance project nodes, and the maintenance project node embedding vector is calculated. On the other hand, based on the second graph attention network, through weighted aggregation, the maintenance stand node aggregates the information of surrounding neighboring nodes, and the maintenance stand node embedding vector is calculated, and then the maintenance stand node embedding vector and the maintenance project node embedding vector are subjected to average pooling operation to obtain the state embedding vector, so that the maintenance stand node embedding vector, the maintenance project node embedding vector and the state embedding vector can be used as the input of the multi-layer perceptron, and the action scalar of each executable OM arc obtained by the multi-layer perceptron processing. Therefore, each type of node is fully extracted separately to avoid interference caused by feature mixing, and at the same time, the mutual influence of the two types of nodes is retained through information transmission between each other. It not only effectively improves the precise expression of node features, but also significantly enhances the flexibility and efficiency in dealing with complex scheduling scenarios, so as to obtain more refined scheduling decisions.

[0092] In one embodiment of the present application, a computer-readable storage medium is provided, on which a program is stored. The stored program includes a method that can be loaded by a processor and process any of the above embodiments.

[0093] Those skilled in the art will appreciate that all or part of the functions of the various methods in the above-mentioned embodiments can be implemented by hardware or by computer programs. When all or part of the functions in the above-mentioned embodiments are implemented by computer programs, the program can be stored in a computer-readable storage medium, and the storage medium can include: read-only memory, random access memory, disk, optical disk, hard disk, etc., and the program is executed by a computer to implement the above-mentioned functions. For example, the program is stored in the memory of the device, and when the program in the memory is executed by the processor, all or part of the above-mentioned functions can be implemented. In addition, when all or part of the functions in the above-mentioned embodiments are implemented by computer programs, the program can also be stored in a storage medium such as a server, another computer, disk, optical disk, flash disk or mobile hard disk, and can be downloaded or copied and saved in the memory of the local device, or the system of the local device is updated, and when the program in the memory is executed by the processor, all or part of the functions in the above-mentioned embodiments can be implemented.

[0094] The above specific examples are used to illustrate the present invention, which is only used to help understand the present invention and is not intended to limit the present invention. For those skilled in the art, according to the concept of the present invention, some simple deductions, modifications or substitutions can be made.

Claims

1. A method for scheduling aircraft maintenance stands based on a heterogeneous graph model, characterized in that: include: Obtain an input feature vector; the input feature vector includes a maintenance station feature vector, a maintenance project feature vector and an OM arc feature vector; wherein the maintenance station feature vector includes: the earliest available time of each maintenance station, the number of maintenance projects that can be repaired by each maintenance station in the current maintenance project, and the utilization rate of each maintenance station; the maintenance project feature vector includes: the maintenance completion status of each maintenance project, the number of available maintenance stations for each maintenance project, the average maintenance time of each maintenance project, the number of maintenance projects that have not started maintenance, and the start maintenance time of each maintenance project; the OM arc feature vector includes the maintenance time of each maintenance project at each available maintenance station; wherein the utilization rate of any maintenance station represents the ratio of the time difference between the start maintenance time of the maintenance project being repaired at any maintenance station and the current time to the time required for the maintenance project being repaired; Constructing a heterogeneous graph model based on the input feature vector, the heterogeneous graph model includes a maintenance stand node and a maintenance item node, wherein, for each maintenance aircraft, a connection arc between maintenance items is constructed according to the dependency relationship between various maintenance items of the maintenance aircraft, and for each maintenance stand, an OM arc is established according to the maintenance items that can be maintained at each maintenance stand; The input feature vector and the heterogeneous graph model are used as inputs of an aircraft maintenance stand scheduling strategy prediction network to achieve prediction of the aircraft maintenance stand scheduling strategy; wherein the aircraft maintenance stand scheduling strategy prediction network is obtained by learning and training in combination with a constructed loss function; Scheduling of aircraft maintenance stands is achieved based on the predicted aircraft maintenance stand scheduling strategy; The method of using the input feature vector and the heterogeneous graph model as inputs of an aircraft maintenance stand scheduling strategy prediction network to achieve prediction of an aircraft maintenance stand scheduling strategy includes: The original feature vector and the heterogeneous graph model are used as inputs. Based on the first graph attention network, through weighted aggregation, the maintenance project node aggregates the information of the neighboring maintenance project nodes, and the maintenance project node embedding vector is calculated. The original feature vector and the heterogeneous graph model are used as input. Based on the second graph attention network, the maintenance station node aggregates the information of the surrounding neighbor nodes through weighted aggregation, and the maintenance station node embedding vector is calculated. Performing a flat pooling operation on the maintenance position node embedding vector and the maintenance item node embedding vector to obtain a state embedding vector; The maintenance position node embedding vector, the maintenance item node embedding vector and the state embedding vector are used as inputs of a multilayer perceptron, and the action scalar of each executable OM arc is obtained through processing by the multilayer perceptron; Each action scalar is normalized to obtain the aircraft maintenance stand scheduling strategy.

2. The aircraft maintenance stand scheduling method according to claim 1, characterized in that: The training method of the aircraft maintenance stand scheduling strategy prediction network includes: Acquire the original feature vectors for training, wherein the original feature vectors include a maintenance position feature vector, a maintenance item feature vector, and an OM arc feature vector; Constructing a heterogeneous graph model based on the original feature vector; The original feature vector and the heterogeneous graph model are used as inputs and combined with the loss function to train the constructed initial architecture for predicting aircraft maintenance stand scheduling strategies, including: The original feature vector and the heterogeneous graph model are used as inputs. Based on the first graph attention network, through weighted aggregation, the maintenance project node aggregates the information of the neighboring maintenance project nodes, and the maintenance project node embedding vector is calculated. The original feature vector and the heterogeneous graph model are used as input. Based on the second graph attention network, the maintenance station node aggregates the information of the surrounding neighbor nodes through weighted aggregation, and the maintenance station node embedding vector is calculated. Performing a flat pooling operation on the maintenance position node embedding vector and the maintenance item node embedding vector to obtain a state embedding vector; The maintenance position node embedding vector, the maintenance item node embedding vector and the state embedding vector are used as inputs of a multilayer perceptron, and the action scalar of each executable OM arc is obtained through processing by the multilayer perceptron; Perform normalization operation on each action scalar to obtain the aircraft maintenance stand scheduling strategy; Using the state embedding vector as an input of a second multilayer perceptron, and obtaining an output as an output scalar after processing by the second multilayer perceptron; A loss function is constructed based on the outputs of the first multi-layer perceptron and the second multi-layer perceptron, and the constructed loss function is back-propagated to the first graph attention network and the second graph attention network to realize the training of the aircraft maintenance stand scheduling strategy prediction network.

3. The aircraft maintenance stand scheduling method according to claim 1 or 2, characterized in that: The first graph attention network includes multiple layers of graph attention networks. Except for the first layer of graph attention network, for any layer of graph attention network, the output of the previous layer of graph attention network is used as the input of the any layer of graph attention network.

4. The aircraft maintenance stand scheduling method according to claim 3, characterized in that: For any maintenance project node, based on any layer of graph attention network, through weighted aggregation, the maintenance project node aggregates the information of neighboring maintenance project nodes, and the maintenance project node embedding vector is calculated, including: Based on the input feature vector and the heterogeneous graph model, the attention coefficients between all maintenance project nodes having connection arcs with the arbitrary maintenance project node are calculated; wherein all maintenance project nodes having connection arcs with the arbitrary maintenance project node include the arbitrary maintenance project node itself; Based on the normalization function, calculate the normalization coefficient of the attention coefficient; Through weighted aggregation, any maintenance project node is aggregated with surrounding neighbor node information, and the maintenance project node embedding vector of the any maintenance project node is calculated.

5. The aircraft maintenance stand scheduling method according to claim 4, characterized in that: The method of calculating the attention coefficients between all maintenance project nodes having connection arcs with any maintenance project node based on the input feature vector and the heterogeneous graph model includes: ； in, Indicates that any maintenance project node O ij There is a connection arc between any maintenance project nodes O ip The attention coefficient of the l-th layer graph attention network, ij represents the j-th maintenance item of the i-th aircraft, ; LeakyReLU represents the LeakyReLU activation function; represents the weight of a single-layer feedforward neural network, and T represents transpose; Represents the calculated maintenance item node O of the l-1th layer graph attention network ij The embedding vector of , when l=1, , Represents maintenance project node O ij The eigenvector of Represents the calculated maintenance item node O of the l-1th layer graph attention network ip The embedding vector of , when l=1, , Represents maintenance project node O ip The eigenvector of express The linear transformation matrix of express The linear transformation matrix of Represents a vector merge operation; The normalization coefficient of the attention coefficient is calculated based on the normalization function, including: ； in, Represents the attention coefficient The normalized coefficient of Represents the self-attention coefficient The normalization coefficient of , softmax represents the normalization function, represents the set of all maintenance project nodes that have connection arcs with any maintenance project node. represents the natural exponential function; The weighted aggregation is used to aggregate the surrounding neighbor node information of any maintenance station node, and the maintenance station node embedding vector of the any maintenance station node is calculated, including: ； in, Represents the maintenance item node O output by the l-th layer graph attention network ij The embedding vector of represents a non-linear activation function.

6. The aircraft maintenance stand scheduling method according to claim 3, characterized in that: The second graph attention network includes a multi-layer graph attention network, which has the same number of layers as the first graph attention network; for the second graph attention network, except for the first layer of graph attention network, for any layer of graph attention network, the output of the previous layer of graph attention network is used as the input of the any layer of graph attention network.

7. The aircraft maintenance stand scheduling method according to claim 6, characterized in that: For any maintenance station node, based on any layer of graph attention network, through weighted aggregation, the maintenance station node aggregates the information of surrounding neighbor nodes, and calculates the maintenance station node embedding vector, including: Based on the input feature vector and the heterogeneous graph model, the self-attention coefficient of any maintenance position node is calculated, and the attention coefficients between all maintenance project nodes that have OM arcs with the any maintenance position node are calculated; Based on the normalization function, the normalization coefficient of the self-attention coefficient and the normalization coefficient of the attention coefficient are calculated; Through weighted aggregation, any maintenance station node aggregates the surrounding neighbor node information, and calculates the maintenance station node embedding vector of the any maintenance station node.

8. The aircraft maintenance stand scheduling method according to claim 7, characterized in that: The method of calculating the self-attention coefficient of any maintenance position node based on the input feature vector and the heterogeneous graph model, and calculating the attention coefficients between all maintenance project nodes having OM arcs with the any maintenance position node, includes: ； ； in, Indicates that any maintenance station node M k There is an OM arc between any maintenance project nodes O ij The attention coefficient of the l-th layer graph attention network, ij represents the j-th maintenance item of the i-th aircraft; Indicates any maintenance station node M k The self-attention coefficient of the l-th layer graph attention network, k represents the k-th maintenance station; LeakyReLU represents the LeakyReLU activation function; represents the weight of a single-layer feedforward neural network, and T represents transpose; represents the calculated embedding vector of the maintenance station node of the l-1th layer graph attention network. When l=1, , represents the maintenance position feature vector; represents the extended feature vector, , Represents the calculated maintenance item node O of the l-1th layer graph attention network ij The embedding vector of , when l=1, , Represents maintenance project node O ij The eigenvector of represents the OM arc eigenvector; express The linear transformation matrix of express The linear transformation matrix of Represents a vector merge operation; The method of calculating the normalization coefficient of the self-attention coefficient and the normalization coefficient of the attention coefficient based on the normalization function includes: ； ； in, Represents the attention coefficient The normalized coefficient of Represents the self-attention coefficient The normalization coefficient of , softmax represents the normalization function, It represents the set of all maintenance project nodes that have OM arcs with any maintenance station node. represents the natural exponential function; The weighted aggregation is used to aggregate the surrounding neighbor node information of any maintenance station node, and the maintenance station node embedding vector of the any maintenance station node is calculated, including: ； in, Represents the maintenance station node M output by the l-th layer graph attention network k The embedding vector of represents a non-linear activation function.

9. The aircraft maintenance stand scheduling method according to claim 1 or 2, characterized in that: The performing of a flat pooling operation on the maintenance position node embedding vector and the maintenance item node embedding vector to obtain a state embedding vector includes: ； in, represents the state embedding vector, Represents the maintenance station node M output by the attention network in the second graph k The embedding vector of Represents a collection of maintenance bay nodes. Indicates the number of maintenance bay nodes. Represents the maintenance item node O output by the attention network of the first graph ij The embedding vector of Represents a collection of maintenance project nodes, Indicates the number of maintenance project nodes, Represents a vector merge operation.

10. The aircraft maintenance stand scheduling method according to claim 9, characterized in that: The maintenance position node embedding vector, maintenance item node embedding vector and state embedding vector are used as inputs of the first multilayer perceptron, and the action scalar of each executable OM arc obtained by processing by the first multilayer perceptron includes: ； in, represents the action scalar, s represents the index of the action scalar, represents a multi-layer perceptron, represents the OM arc eigenvector; The normalization operation for each action scalar includes: ； in, Represents action scalar The probability of selection, represents the natural exponential function.

Images